The second city has demand, a cheaper cost base and a person who wants to run it. The model shows the office paying for itself in eleven months.
The model is about cost. What actually changes is the number of distinct situations arriving at people whose set of available responses stays exactly the same.
Why these three models
The decision is whether to open a second location, and what would have to be true first. The features that fire are a control problem that grows faster than the headcount, feedback loops that get longer in both directions, and a commitment that is far more expensive to reverse than to make.
Three lenses. Variety produces an equilibrium answer about whether two sites can be regulated at all. Delay produces a complex answer about why your corrections will overshoot. Option value produces a random answer about why the lease term matters more than the rent. The first is the one that decides whether this works and the one nobody runs, because it does not look like a financial question.
1. Count the situations, then count the responses
A regulator can only control a system if it commands at least as many distinguishable responses as the system has distinguishable states. Fewer responses than states means the system cannot be regulated, no matter how hard anyone works or how many people you add.1
Apply that to a second site honestly. One office has one set of local rules, one regulator, one bank, one labour market, one landlord, one internet provider, one set of public holidays. Two offices do not have twice as many of each; they have those pairs plus every situation that arises from the interaction, which is where the growth actually comes from. Which site owns this customer. Whose margin does the transfer sit in. Which handbook applies to someone who moved.
Now count the responses. If your management response set is the same as it was, you have added states without adding variety, and the shortfall shows up as escalation: everything unusual reaches the founder, because the founder is the only response category that covers unusual.
There are exactly two repairs and hiring is not reliably either of them. Attenuate the incoming variety: standardise the offer so the second site sells the same three things, restrict what it may decide locally, use one contract template. Or amplify your response set: give the site head a real decision boundary with a budget attached, so a whole class of situation is handled without reaching you.
Adding a person with no new response category changes neither count. You cannot fix a shortfall of variety with a shortfall of people.
Build the list rather than estimating it, because the count is always higher than the estimate. Write one row for each decision the second site will actually face in a month: approving a discount, hiring a driver, paying a supplier early, replacing a failed generator, extending credit to a customer who asks, opening on a local holiday the other city does not observe. Next to each, write who decides today and what it costs when the answer is wrong. Twenty rows is normal. Delegate first the rows where the cost is small and the answer is you, because those are the ones that arrive weekly.
Attenuation carries a cost people discover late. Standardising the offer down to three things is the fastest way to make two sites governable, and it can also remove the reason the second city looked attractive. If that market wanted a different payment structure, a different pack size or a different service level, standardising it away leaves you with an office selling a product the city did not ask for. So attenuate the internal variety, meaning contracts, approvals, systems and reporting, and leave the offer variable only where you have already seen demand. The test is whether the difference was requested by a paying customer or proposed by the person who wants the office.
2. Every loop between the sites gets longer
The second lens explains why the corrections you make will be badly sized.
In one office, the gap between something going wrong and you knowing about it is short, often minutes. Across two, it lengthens: the information has to be noticed by someone local, judged worth reporting, reported, and read. Each stage adds delay, and the delay applies in both directions, so your instruction also arrives late.
A control loop with a long delay does not merely respond slowly, it overshoots. You keep pushing because nothing has come back, the accumulated effect arrives, you correct hard, and the correction itself arrives late.2 The pattern is recognisable in any multi-site company: a quarter of under-management followed by a visit, a reorganisation, and a period of over-management that the site experiences as arbitrary.
The remedy is not to travel more. It is to shorten the measurement rather than the response: pick three observables the second site reports weekly, unfiltered, on a fixed schedule, chosen because they move early rather than because they matter most. Cash collected, not revenue booked. Escalations raised, not problems solved. Days since the site head last spoke to a customer.
It is worth knowing how robust this failure is before you assume good people will avoid it. In a controlled four-layer chain with demand held nearly constant, participants including experienced executives produce large oscillations anyway, run average costs about ten times the optimum available from the information they actually hold, and afterwards describe an external cause that did not occur.4 Nobody in that experiment was careless, and every one of them blamed the weather.
Choose those three so one person can produce them in ten minutes, or they will not survive a busy month. None of them should need a system to exist first, and each should be a number the site head already knows without looking.
Unfiltered reporting is something you pay for, not something you request. A site head who reports a bad week and receives a call from head office learns in one cycle to report the bad week later, and your delay has just grown by exactly the amount you were trying to remove. So meet the first bad report with a question and no instruction, deliberately and visibly. The tell that you have already lost this is that the numbers arrive smooth. Real weekly operating figures are lumpy, and a series with no bad weeks in it is a report about the reporter.
And size corrections as though you are seeing an old picture, because you are. The instinct to make one large adjustment on a visit is exactly the input that produces the swing. Measure the actual lag once: pick three things that went wrong and count the days between the event and your hearing about it. If the answer is eleven days, then everything on your dashboard is eleven days old and somebody has already responded to it locally. Make what has changed since this number was produced the first question on every site call.
3. The lease is the decision, not the rent
The third lens moves the question from whether to how, and it is the cheapest thing in this article to act on.
Opening is easy and closing is not. Closing a site means redundancies, a landlord negotiation, a story your market will interpret, and a team who watched it happen. The value of keeping a decision open rises with how uncertain you are and with how expensive the reversal would be, and both are high here.3
So buy the reversibility rather than the space. Serviced offices at a premium instead of a three-year lease. Two people on contracts before four on payroll. A local partner arrangement before an entity. Each costs more per month and each converts an irreversible commitment into a reversible one, which is the same as manufacturing an option you did not have.
The premium looks like waste in the twelve months where nothing goes wrong, which is why finance-minded teams remove exactly these line items. Their absence is discovered in the quarter you need to change direction, and by then the option cannot be bought at any price.
The cheap versions are worth listing, because a serviced office does not exist in every city. Sublet a room from a business already there, month to month. Take desks in a shared workspace. Put two people in the field with a laptop and no premises for a quarter. Trade volume for counter space at a distributor who is already selling to your buyers. Each of those buys the same information the office would, and none of them commits you to a term.
Where a real lease is unavoidable, the terms worth negotiating are not the ones people negotiate. Rent is the smallest number on the page. Ask for a break at twelve months, a deposit of two months rather than six, and a right to sublet with consent that cannot be unreasonably withheld. A deposit sitting with a landlord for three years is working capital you have lent at no interest, and where recovering it is slow it should be carried as a cost rather than as an asset.
Then set the trigger before you sign: the observable that means this is not working and you will close it, chosen now while nothing is at stake. Write it so somebody else can check it. If the site has not covered its own direct costs by month nine, or if the site head leaves and is not replaced within eight weeks, it closes. Name who checks and on what date. By month nine the site will have a team, a customer list and a story about the quarter it is about to have, and none of that is evidence. A site that has never had a written closing condition does not get closed, it gets defended.
What the three say together
- Count the distinguishable situations two sites will generate, and the responses you have. If responses are fewer, standardise the offer or delegate a real decision boundary before you open.
- Choose three early observables the second site reports weekly and unfiltered. Measure early, not important.
- Buy reversibility: serviced space, contracts, a partner. Pay the premium in the good months.
- Write the closing trigger before you sign the lease.
Where they disagree
The variety argument and the option argument recommend different sequences.
Variety says do the work first: standardise, delegate, build the response set, and only then open, because a site opened into a control shortfall generates problems faster than you can build the machinery to handle them. The option argument says that all of that is expensive preparation for a decision you might reverse, and that the cheapest way to learn whether a second site works is a small reversible version of it, run before the standardisation exists.
The resolution is that they are right about different failure modes. The reversible pilot is correct for testing demand, which is the question most founders think they are asking. The variety work is correct before scale, because a pilot that succeeds will be expanded, and the expansion is where the control shortfall becomes visible. Run the pilot without the standardisation, then do the variety work before the second site becomes permanent, and do not let a successful pilot skip that step because it appeared to work at three people.
What none of them contain
None of the three prices what a second site does to the first. Attention moves, the founder is visibly elsewhere, and the original team reads that as a statement about where the future is. That cost is real, it lands immediately, and none of these lenses sees it.
None of them handles the site opened for a reason other than the market: a co-founder relocating, a grant condition, an investor’s preference. Those are common and the analysis here will produce a confident answer to a question that was never the real one.
None of them handles the second site that is genuinely a different business. A city with different buyers, different regulation and a different sales motion is not an extension of your company, it is a second company sharing a brand, and every efficiency in the plan assumed otherwise. That distinction is usually visible in the first quarter of revenue mix and almost never revisited afterwards.
And one property the ensemble will not produce: the two sites will diverge in culture whatever you do. Distance plus local hiring plus different customers produces different norms, and the models treat the company as one system being extended when it is closer to two systems being coupled. The interesting question is which differences you want to permit, and it is a question none of this arithmetic will answer for you.
The one action that survives the ignorance: before signing anything, write the list of decisions the second site will need to make locally, and next to each, who decides today. Every row where the answer is you is a situation with no response attached, and the count of those rows is the real readiness measure.
Who has to move
The founder, and the work belongs before the lease rather than after the opening, because after the opening every problem is urgent and the standardisation never gets built. The instinct is to hire a strong site head and delegate, which is correct only if that person is given a decision boundary rather than a title; without one they become another escalation path. The cheapest first test is the decision list, which takes an hour and typically shows that a dozen ordinary local calls currently route through one person.
Sources and notes
- Stafford Beer, Cybernetics and Management, English Universities Press, 1959. The law of requisite variety, attributed to Ashby, is adopted in the chapter on the Black Box as an axiom in the process of controlling complex systems, with the condition stated as sufficient permutative variety to provide a one-to-one transformation from the controlled system into the control box. The two repairs of attenuating incoming variety and amplifying the regulator’s are named as such in Brain of the Firm, Allen Lane, 1972. Used throughout section 1. Note on attribution, since it is commonly got wrong: the Viable System Model and the Systems One through Five architecture belong to Brain of the Firm in 1972, not to this book.
- John D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill, and Donella H. Meadows, Thinking in Systems: A Primer, Chelsea Green Publishing. The behaviour of a balancing loop containing a delay, including overshoot and oscillation when the correction is sized as though the delay were zero, is developed in both. Used in section 2. The prescription that follows, to measure closer to where the state changes rather than to respond faster, is the standard remedy.
- Sanjit S. Dhami, The Foundations of Behavioral Economic Analysis, Oxford University Press, and the real-options literature it references. The property used in section 3 is structural: the value of holding a decision open rises with the irreversibility of the commitment and with the uncertainty to be resolved. No option value is computed, since the inputs required for a valuation are not available for a site-opening decision.
- John D. Sterman, The Beer Distribution Game, MIT, https://web.mit.edu/jsterman/www/SDG/beergame.html. Participants with good local information and poor global information generate large oscillations from a stable environment, at average costs about ten times the optimum available from what they actually hold, and afterwards attribute the swings to external causes that did not occur. Cited in section 2 as evidence that the multi-site swing is structural rather than a failure of the people running each site.
A note on a number this article does not give. There is no revenue level at which a second site becomes correct. It depends on how standardised your offer is and how much of your management response set is delegated, both of which you can measure and neither of which appears in a benchmark. The decision list is the measurement.
Joshua Agonya Pi’Rwot, Founder.